MU-GeNeRF

This repository contains a trained checkpoint for MU-GeNeRF: Multi-view Uncertainty-guided Generalizable Neural Radiance Fields for Distractor-aware Scene.

MU-GeNeRF is a generalizable neural radiance field model that uses multi-view uncertainty modeling for distractor-aware novel-view synthesis.

Files

• model.ckpt: PyTorch Lightning checkpoint from epoch 31 of a 32-epoch run.
• config.yaml: Training and model configuration used for the checkpoint.
• requirements.txt: Python dependencies for the original implementation.

The checkpoint is a custom PyTorch Lightning checkpoint. It is not directly loadable with the Transformers or Diffusers pipelines.

Usage

The original implementation is available at:

https://github.com/Yanyilucas/MU-GeNeRF

Clone the project, install its dependencies, and place model.ckpt at ckpts/epoch-epoch=31.ckpt. Evaluation is then run with:

./eval_test.sh

The evaluation setup requires a compatible CUDA/PyTorch environment, the configured datasets, and precomputed DINOv2 ViT-S/14 patch features under each scene's dino_feature/ directory. The model configuration uses 360x640 input images, four source views during training, and eight source views during test evaluation.

Training configuration

The checkpoint was trained with the DF3DV-1K-Star and On-the-go dataset configuration. RobustNeRF was disabled in the saved training configuration. The model uses 384-dimensional DINOv2 features and uncertainty-aware losses.

Limitations

This release contains weights and configuration only. Dataset files and precomputed feature caches are not included. Dataset paths in the original training configuration are machine-specific and should be adjusted for a new environment.

License

No license is specified for this model release. Please refer to the source project and its authors before redistribution or commercial use.